Summaries now come back in the language you expect.

Fixed — output language was unpredictable (#16)
Every prompt is written in English and nothing stated which language to answer in, so the model inferred one. A mostly-Chinese document carrying a few English acronyms would come back as a long English summary, while a pure-Chinese one stayed Chinese. Every request now states the language explicitly.

New — choose the summary language
A Summary language picker on the Config screen and -l/--language on the CLI. The default, Auto, matches the document. Any language name is accepted, not only the listed ones.

docsummarizer-cli notes.txt -l Chinese

Fixed — release binaries were built without Qt
The v2.0.0 downloadable assets were built without the gui extra, so they failed at startup with ModuleNotFoundError: No module named PySide6. Anyone who hit that: the assets below are the working builds.

Windows and Linux builds are CPU-only; the ~2.5 GB model downloads on first run.

DocSummarizer 2.0.0

A ground-up redesign of the desktop application around a new offline-first interface.

Highlights

• New desktop UI — a redesigned two-pane console: brief / detailed / structured summaries, a raw-extraction view with search, and batch processing with live per-file status.
• Source-grounded provenance — click a summary point to highlight the exact sentence it was drawn from.
• Cancellable summaries — Stop interrupts a running summary immediately.
• Whole-document summarization — long documents are processed in chunks (map-reduce), not truncated.
• Refined visual design with bundled typography.
• Accurate compute reporting — the GPU toggle reflects what the build actually supports.

Notes

• The portable DocSummarizer.exe runs entirely on CPU and needs no GPU or installation.
• On first launch it downloads the local model (~2.5 GB, one-time); everything runs offline afterwards.

Windows portable executable attached below.

Security

• Bump llama-cpp-python to the 0.3.x line (clears GHSA-56xg-wfcc-g829, a critical RCE via Jinja2 SSTI in chat-template parsing).
• Bump pypdf to the 6.x line (clears ~24 DoS / RAM-exhaustion advisories that apply to attacker-supplied PDFs).
• Bump tqdm past CVE-2024-34062 (CLI eval injection; not exploitable in this app but bumped on principle).
• Move every other dependency to its current major.

Reliability

• Fix Tk widget mutation from worker threads across the load, download, summarize, and batch paths.
• Lock-protected Summarizer access with explicit close() + context-manager support (replaces unreliable __del__).
• Fix PyInstaller spec on macOS — it was silently bundling zero llama backend libraries because the lookup path was hardcoded to venv/ / venv_linux/.
• Drop the .doc extension claim (legacy OLE compound; python-docx cannot read it). .docx only.
• Single stat() in document_parser metadata to avoid a network-share race.

UX

• Real progress bar during the 4.4 GB model download (was 0% → 100% only).
• Initial model load now shows a loading dialog and disables inputs.
• Batch processing reports per-file success/failure with a list of failures in the completion popup.

Infrastructure

• The package is now installable: src/docsummarizer/ layout, pip install -e ., docsummarizer and docsummarizer-cli entry points.
• llama-cpp-python lives behind the [runtime] extra so CI installs the package without the heavy native backend.
• GitHub Actions CI on Ubuntu × Python 3.10 / 3.11 / 3.12 — ruff check, ruff format check, 42 pytest tests.
• Hash-pinned requirements.lock (generated by uv pip compile --generate-hashes --all-extras) for reproducible release builds.
• Dependabot enabled (weekly, pip + github-actions).
• Branch protection on master requiring all three matrix CI jobs.

Upgrade notes

The package install path changed: pip install -r requirements.txt no longer works (the file is gone). Use the setup scripts (setup_and_run.{sh,bat}) — they now run pip install -e ".[runtime]". For development, pip install -e ".[dev]" is enough (skips the C++ backend; tests stub it).

What's New

CPU Usage Control

• Added CPU threads slider in Settings tab (2 to max cores)
• Users can now balance between speed and system responsiveness
• Default: uses half of available CPU cores instead of all

Improved User Experience

• Added loading dialog with progress indicator during model reload
• Visual feedback when changing thread settings

Logging & Diagnostics

• Added logging system for troubleshooting
• Logs stored in %LOCALAPPDATA%\DocSummarizer\logs\ (Windows) or ~/.local/share/DocSummarizer/logs/ (Linux)
• Tracks: startup info, model loading time, summarization speed (tokens/sec)
• No sensitive data logged (document content is never stored)

Bug Fixes

• Fixed "NoneType object has no attribute 'write'" error when downloading model on Windows
• Fixed "system cannot find the path" error when loading model (llama_cpp DLLs now bundled)
• Fixed window not closing properly on X button
• Reduced CPU usage during inference with n_threads_batch optimization